mksethi/eli5_sae_features

Dataset Card for gpt2eli5sae_features

This dataset aims to create a corpus of data to help guide research into monosemantic features using SAE's. It has been generated using this raw template

Dataset Details

Dataset Description

This dataset takes the eli5 subset from facebook/kilt_tasks and processes it to be used in SAE research. More specifically, we take the inputs, and tokenize them using the gpt2-small tokenizer. The outputs are tokenized, embedded , and encoded using the

• —Curated by: Manik Sethi
• —License: MIT

Dataset Sources [optional]

<!-- Provide the basic links for the dataset. -->

• —Repository: hallucination_circuits
• —Paper: On the way!
• —Demo: On the way!

Uses

<!-- Address questions around how the dataset is intended to be used. -->

Direct Use

<!-- This section describes suitable use cases for the dataset. -->

This dataset is meant to train a multi-layer-perceptron to predict what the SAE feature activations would be in the answer for a prompt. Note that this dataset doesn't actually have the answers to the tokenized questions, but rather a ~25k long vector of activations for different features in a trained SAE.

Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

Currently, we only have the train split from the ELI5 dataset. Therefore, this dataset is very similarly structured.

Dataset Creation

Curation Rationale

<!-- Motivation for the creation of this dataset. -->

[More Information Needed]

Source Data

The source of this data is from the subreddit "Explain Like I'm 5", or ELI5 for short.

Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

The inputs are tokenized prompts. The labels are tokenized, embedded, and then encoded with the given SAE layer.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

[More Information Needed]

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.

Dataset Card Contact

mksethi@ucdavis.edu